Short answer
Nobody knows for certain how to build superintelligence (SI), an intellect that greatly exceeds the best humans in virtually every domain. But the plausible routes can be mapped. The leading pathway today is scaling general AI models plus reinforcement learning, leading to AI systems that automate AI research itself, which could trigger rapid self-improvement. Alternative routes include brain emulation, neuromorphic hardware, collective AI systems and human-AI merger. Every route runs into the same six bottlenecks: compute, energy, data, algorithms, real-world experimentation and alignment. This guide maps each pathway, rates the evidence and lists the signals that would show SI is approaching.
What "achieving superintelligence" actually means
Before asking how, we need a usable target. Philosopher Nick Bostrom distinguished three forms of superintelligence, a framework explained in our review of his book:
| Form | What it means | Example |
|---|---|---|
| Speed superintelligence | Thinks like a smart human, but much faster | A human-level mind running 1,000× faster |
| Collective superintelligence | Many minds coordinated better than any human organisation | Millions of AI researchers working in parallel |
| Quality superintelligence | Thinks in ways qualitatively smarter than humans | Insights no number of humans would find |
A practical, testable definition for this article: a system or coordinated group of systems that outperforms the best human experts and the best human teams at nearly all cognitive work, including science, engineering, strategy and AI research itself. That last clause matters, because an AI that is better than humans at improving AI changes the speed of everything afterwards. See AGI vs ASI for how this differs from human-level AGI.
The eight pathways to superintelligence
Pathway 1: Scaling foundation models
The idea: keep increasing compute, data and model size. Scaling laws show that model loss falls predictably as training compute grows, and training compute for frontier models has grown roughly four to five times per year for a decade, according to Epoch AI.
Evidence for: each order of magnitude of compute has brought new abilities, from coherent essays to graduate-level science answers and competition maths. Rich Sutton's influential essay *The Bitter Lesson* argues that general methods which leverage computation have consistently beaten hand-crafted approaches over 70 years of AI research.
Evidence against: high-quality human text is finite, pre-training gains per dollar diminish under power laws, and some researchers argue next-token prediction alone lacks the grounding and planning needed for superhuman reasoning.
Assessment: necessary but probably not sufficient on its own. Scaling provides the raw capability; other pathways turn it into superhuman performance.
Pathway 2: Reinforcement learning and test-time reasoning
The idea: train models with reinforcement learning on problems with verifiable answers, and let them think longer at inference time. This is how reasoning models reached top human levels in competition maths and coding.
Why it matters for SI: RL is the only known method that has produced clearly superhuman performance in open-ended domains. AlphaGo's famous "move 37" against Lee Sedol in 2016 was a strategy human experts had not considered. Self-play let AlphaZero surpass all human chess knowledge in hours.
The open question: can RL work beyond domains with clean, automatic rewards (maths, code, games) and generalise to messy domains like science, strategy and judgement, where success is slow or ambiguous to measure?
Pathway 3: Automated AI research (the recursive pathway)
The idea: build AI systems good enough to do AI research: designing architectures, writing training code, running experiments and analysing results. Then use them to build better AI, which builds better AI again. This is the core of the intelligence explosion hypothesis first proposed by mathematician I. J. Good in 1965.
Why it is the most discussed pathway: AI research is digital work done at computers, exactly what AI agents are getting better at. Early signs already exist: Google DeepMind's AlphaEvolve (2025) discovered a way to multiply 4×4 complex-valued matrices with 48 scalar multiplications, improving on a result that had stood since 1969, and found optimisations used in Google's own data centers and chip design.
The key uncertainty: whether research progress is limited mainly by thinking (which AI can multiply) or by compute for experiments (which grows only as fast as hardware). If compute is the bottleneck, millions of AI researchers might speed progress only modestly. The scenario forecast AI 2027 explores the fast version of this pathway.
Pathway 4: Agents that learn from the real world
The idea: today's models learn mostly from static human data. Agents that act in the world, running experiments, using tools, controlling robots and learning continuously from experience, could gather knowledge humans do not have. Researchers David Silver and Richard Sutton describe this as moving from the "era of human data" to an "era of experience".
Bottleneck: the physical world is slow. A chemistry experiment takes hours and a clinical trial years, no matter how smart the planner. Simulation, automated laboratories and parallel robotics are the levers.
Pathway 5: Whole brain emulation
The idea: scan a human brain at sufficient resolution and simulate it on computers. Run it faster, copy it, and improve it.
Progress: in 2024 the FlyWire project published the complete wiring diagram of an adult fruit fly brain, about 140,000 neurons. In 2025 the MICrONS project mapped a cubic millimetre of mouse visual cortex, with around 200,000 cells and more than 500 million synapses. A human brain has about 86 billion neurons and on the order of 100 trillion synapses.
Assessment: a slow, data-heavy pathway with very large scanning, storage and modelling gaps. It is unlikely to arrive before machine-learning approaches, but it remains the only route that starts from a known working general intelligence.
Pathway 6: Neuromorphic and novel computing
The idea: build hardware that works more like brains, with spiking neurons, in-memory computing or analogue chips, to escape the energy limits of today's GPUs. The human brain runs on about 20 watts, while a single frontier training cluster can draw tens to hundreds of megawatts. See how much energy AI uses.
Assessment: more likely to make superintelligence cheaper and more widespread than to be the first route to it. Optical computing and quantum computing are sometimes mentioned, but there is no current evidence that quantum computers speed up general learning.
Pathway 7: Collective superintelligence
The idea: superintelligence might not be one giant model but millions of coordinated AI instances, plus tools, databases and humans, forming an organisation smarter than any company or government. A frontier model can already be run as many parallel copies; the hard parts are coordination, shared memory and error correction.
Why it matters: this pathway could produce superhuman results without any single system being superhuman, which makes it harder to notice and govern.
Pathway 8: Human-AI merger and biological enhancement
The idea: brain-computer interfaces, cognitive enhancement or genetic selection could raise human intelligence, or merge humans with AI. Brain implants already let people with paralysis control computers.
Assessment: current interfaces transmit far less information than speech or typing. This pathway is unlikely to keep humans competitive with machine intelligence on its own, but it could shape how people relate to and oversee SI.
Pathway comparison
| Pathway | Evidence today | Likely speed | Main bottleneck | Contribution to first SI |
|---|---|---|---|---|
| 1. Scaling | Strong | Steady | Compute, data, energy | Foundation |
| 2. RL and reasoning | Strong in verifiable domains | Fast | Reward design outside maths and code | Very high |
| 3. Automated AI research | Early | Potentially very fast | Experiment compute | Very high |
| 4. Real-world agents | Early | Slow to medium | Physical experiment speed | Medium to high |
| 5. Brain emulation | Weak | Slow | Scanning and modelling | Low |
| 6. Neuromorphic and novel hardware | Weak to medium | Slow | Manufacturing, software | Low (efficiency later) |
| 7. Collective SI | Medium | Medium to fast | Coordination, reliability | High |
| 8. Human enhancement | Weak | Slow | Biology, bandwidth | Low |
The most plausible route combines pathways 1, 2, 3 and 7: large foundation models, trained heavily with RL, deployed as millions of coordinated agents that accelerate AI research.
The six bottlenecks
1. Compute
Training runs now use tens to hundreds of thousands of GPUs. Supply depends on chip fabrication, advanced packaging and high-bandwidth memory, all of which have been tight. See What is an AI data center?.
2. Energy
The IEA projects data center electricity demand roughly doubling to about 945 TWh by 2030. Gigawatt-scale AI campuses need new power plants and transmission lines, which take years to build.
3. Data
Epoch AI estimates the stock of public human text could be fully used by frontier training around the late 2020s. Solutions include synthetic data, RL on self-generated problems, multimodal data (video, robotics, sensors) and learning from experience.
4. Algorithms
Algorithmic progress has roughly halved the compute needed for a given level of language-model performance every eight to nine months, according to Epoch AI research. Breakthroughs like transformers or RL reasoning can shift the curve suddenly, which makes timelines hard to predict.
5. Real-world experimentation
A superintelligent scientist is still limited by how fast experiments run. Automated labs, high-fidelity simulation and parallel robotics determine how quickly intelligence turns into discoveries.
6. Alignment and control
Building a system smarter than its creators safely is an unsolved problem. Safety work may slow deployment, and serious accidents could trigger regulation or pauses. See What is AI alignment? and the superintelligence statement calling for a prohibition on developing SI until it can be done safely and with public support.
A staged roadmap
| Stage | Capability | Signals |
|---|---|---|
| 0. Today | Expert-level answers, strong coding, limited autonomy | Top scores on exams and competitions |
| 1. Autonomous agents | AI completes multi-day professional tasks reliably | Agents doing weeks of real work with little supervision |
| 2. AI research assistants | AI substantially speeds up AI labs' own research | Labs reporting large R&D speed-ups from internal AI |
| 3. Automated AI researcher | AI does most AI research end to end | Novel architectures and training methods designed mainly by AI |
| 4. Broad superhuman science | AI leads discoveries across many fields | Rapid, verified breakthroughs in biology, materials, physics |
| 5. Superintelligence | Outperforms the best human teams at nearly everything | Humans can no longer meaningfully evaluate frontier AI work |
The SI proximity dashboard: signals to watch
1. Task horizon: the length of tasks AI agents can complete reliably. Research group METR found this horizon had been doubling roughly every seven months. 2. AI-written AI: the share of frontier lab code, experiments and papers produced by AI. 3. Original discoveries: verified new results in maths and science credited mainly to AI. 4. Compute and power: announcements of gigawatt-scale clusters and next-generation chips such as NVIDIA Rubin. 5. Benchmark saturation: the hardest tests, such as Humanity's Last Exam and FrontierMath, approaching expert ceilings. See AI benchmarks explained. 6. Autonomous replication and research ability in lab safety evaluations. 7. Economic signals: fast productivity growth in software, science and services that cannot be explained otherwise. 8. Lab statements: organisations such as OpenAI, Google DeepMind, Anthropic, Meta Superintelligence Labs and Safe Superintelligence Inc. increasingly describe SI as an explicit goal.
When could it happen?
Forecasts vary enormously. A 2023 survey of 2,778 AI researchers produced an aggregate 50% chance that unaided machines would outperform humans at every task by 2047, while some lab leaders predict systems smarter than humans in most areas within a few years. Others think fundamental breakthroughs are still missing. Our article When will AGI arrive? compares forecasts in detail. The honest answer: the timeline depends mostly on whether automated AI research works, and that is not yet known.
Frequently asked questions
Is superintelligence possible at all?
There is no known physical law that prevents it. Human brains are made of ordinary matter, run on 20 watts and evolved under severe constraints such as skull size. Machines face different limits and can be copied, sped up and scaled. Whether we can build SI soon is uncertain; whether it is possible in principle is widely accepted among researchers.
Will superintelligence come from one company?
It could come from one lab, several labs at once, or a mix of open and closed systems. A single leader is more likely if self-improvement is very fast; a multipolar outcome is more likely if progress is gradual. See what happens after superintelligence.
Does superintelligence need consciousness?
No. Intelligence means the ability to achieve goals and solve problems. Whether advanced AI could be conscious is a separate, open philosophical question.
Can we stop or slow it?
Policy proposals include international compute governance, licensing of frontier training runs and binding safety standards. Whether they are adopted is a political question as much as a technical one.
Sources and further reading
- Bostrom, *Superintelligence: Paths, Dangers, Strategies* (Oxford University Press, 2014)
- Good, I. J., "Speculations Concerning the First Ultraintelligent Machine", *Advances in Computers*, vol. 6 (1965)
- Sutton, The Bitter Lesson (2019)
- Google DeepMind, AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms (2025)
- Grace et al., Thousands of AI Authors on the Future of AI (2024)
- METR, Measuring AI Ability to Complete Long Tasks (2025)
- Dorkenwald et al., Neuronal wiring diagram of an adult brain (FlyWire, Nature, 2024)
- Epoch AI, Trends in machine learning




Leave a Reply